GLM-5.3 open weights: 753 billion parameters
Z.ai's coding model is on Hugging Face with 753 billion parameters and up to 1 million tokens of context, but license terms stay unstated.
Illustration: a data center aisle at night, long rows of servers and cool indicator light on a polished concrete floor.
Z.ai has released GLM-5.3 as an open-weights model on Hugging Face, a 753-billion-parameter mixture-of-experts system downloadable in Safetensors format.
At a glance
- 753 billion parameters, mixture-of-experts, tensor types BF16, F8_E4M3 and F32.
- Maximum context length listed on the model card: 1 million tokens.
- 27 derived quantizations, including builds for llama.cpp, LM Studio, Jan and Ollama.
- CyberGym: 84.5 for GLM-5.3 against 83.3 for DeepSeek-V4 Pro — vendor-reported.
- The license is named glm-5.3; its actual terms do not appear on the model page.
Z.ai has published GLM-5.3 with open weights on Hugging Face. The model card lists 753 billion parameters and a mixture-of-experts design. Files ship in Safetensors format, in BF16, F8_E4M3 and F32 tensor types. Maximum context length is given as 1 million tokens.
What the model card documents
The license carries the name glm-5.3. Its terms are not printed on the page used as the source here. So “open weights” describes the availability of the files, not what you are permitted to do with them. Anyone planning commercial use should read the license text first.
For serving, the card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth and Ascend NPU platforms. Twenty-seven derived quantizations are indexed, among them builds for llama.cpp, LM Studio, Jan and Ollama. Together AI appears as an inference provider. The page reports 8,804 downloads for the month before retrieval.
The benchmarks are vendor numbers
The card places GLM-5.3 against several stronger models. Every figure below comes from Z.ai and was not reproduced for this article.
| Test | GLM-5.3 | Comparison |
|---|---|---|
| Terminal Bench 2.1 | 88.2 | 88.8 (GPT-5.6 Sol) |
| Terminal Bench 3.0 | 28.3 | 34.6 (GPT-5.6 Sol) |
| DeepSWE (v1.1) | 66.9 | 72.7 (GPT-5.6 Sol) |
| CyberGym | 84.5 | 83.3 (DeepSeek-V4 Pro) |
| ExploitBench | 54.4 | 78.0 (GPT-5.6 Sol) |
GLM-5.3 trails GPT-5.6 Sol on three of the five listed tests and leads DeepSeek-V4 Pro on CyberGym. The widest gap is ExploitBench, 54.4 against 78.0. Z.ai claims the top open-model result on Terminal Bench 3.0 and on Agents' Last Exam.
Security-adjacent capability
Z.ai describes an emergent cyber capability and a best-among-open-models result on CyberGym, a test for finding software vulnerabilities. Against GLM-5.2 the company reports a 50 percent gain on its in-house Z.ai Code Bench, from the same base model plus additional post-training. A technical report is linked under arXiv identifier 2602.15763, dated February 17, 2026.
A model distributed openly that also leads on vulnerability discovery is available to defenders and attackers alike. What safeguards Z.ai applies around that is not addressed on the model page.
What stays unverified
Exactly one independent source was available for this report. That leaves several points unchecked: active parameters per token, the wording of the license, any API pricing, and whether third parties have reproduced the benchmark scores. The release date of the weights could not be confirmed separately from the date of the technical report.
FAQ
How many parameters does GLM-5.3 have?
The model card lists 753 billion parameters in a mixture-of-experts architecture. How many are active per token is not stated on the page.
Can GLM-5.3 be used commercially?
The model page does not answer that. It names the license glm-5.3 without printing its terms, so read the license text before deploying anything.
How can I run GLM-5.3 myself?
The card lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth and Ascend NPU, plus 27 quantizations for tools such as llama.cpp and Ollama.
